MS-COCO + Automotive Gen1 Datasets
收藏资源简介：
Spiking neural networks (SNNs) are emerging as a promising evolution in neural network paradigms, offering an alternative to conventional convolutional neural networks (CNNs). One of the most effective methods for SNN development is the CNN-to-SNN conversion process. However, existing conversion techniques are hindered by long temporal durations, or inference latencies, which negatively impact the accuracy of the converted networks. Additionally, the application of SNNs in object detection tasks remains largely under-explored. In this study, we propose a novel approach utilizing a bistable integrate-and-fire (BIF) neuron model integrated with a single-shot multibox detector (SSD) as the detection head. Leveraging the proposed BIF neuron framework, we convert the widely used ResNet architecture into an SNN. We validate the effectiveness of our approach through object detection tasks on the MS-COCO and Automotive GEN1 datasets. Experimental results show that our conversion technique facilitates object detection with reduced temporal steps and significant enhancements in mean average precision (mAP), achieving mAP@0.5 scores of 0.476 and 0.591 for the MS-COCO and Automotive GEN1 datasets, respectively. This research marks the first application of BIF neurons to object detection, presenting a novel advancement in the field.
脉冲神经网络（Spiking Neural Networks, SNNs）作为神经网络范式中极具前景的演进方向，为传统卷积神经网络（Convolutional Neural Networks, CNNs）提供了一种替代方案。当前SNN开发中最有效的方法之一便是CNN转SNN的转换流程。然而，现有转换技术受限于过长的时间跨度或推理延迟，会对转换后网络的精度产生负面影响。此外，SNN在目标检测任务中的应用仍未得到充分探索。 本研究提出一种全新方法，将双稳态积分-放电（Bistable Integrate-and-Fire, BIF）神经元模型与作为检测头的单步多框检测器（Single-Shot Multibox Detector, SSD）相结合。借助所提出的BIF神经元框架，我们将当前广泛使用的ResNet架构转换为SNN。我们通过在MS-COCO与Automotive GEN1数据集上开展目标检测任务，验证了所提方法的有效性。实验结果表明，我们的转换技术可在减少时间步长的同时，显著提升平均精度均值（Mean Average Precision, mAP），在MS-COCO与Automotive GEN1数据集上分别实现了0.476与0.591的mAP@0.5得分。本研究首次将BIF神经元应用于目标检测领域，为该领域带来了全新的技术进展。
